Assessing a RISCy business.

Albert Einstein once remarked that things should be made as simple as possible, but no simpler. Computer architects are beginning to learn a similar lesson when deciding how a computer should translate commands written in a language like FORTRAN into electrical signals that ripple through an integrated-circuit chip. Instead of packing microprocessors with circuits for more, increasingly complex instructions, some designers are "wiring in" only those for a few, frequently used commands. The result is a new class of streamlined computers called "reduced instruction set computers" or RISC machines.

A simple analogy illustrates the idea. The instructions that tell a computer chip how to shift "bits" (electrical pulses) while doing elementary operations such as LOAD, STORE, COMPARE and ADD are somewhat like the keys on a calculator keyboard. Some calculators feature dozens of keys, many of which perform special functions like taking square roots or calculating standard deviations. These calculators tend to be expensive, many keys are rarely used, and users face a bewildering array of choices when trying to solve a problem. In contrast, simple calculators stick to the basics: addition, subtraction, multiplication, division and little more. For many calculations, this is enough.

In a RISC machine, a small set of built-in instructions allows plenty of extra room on a chip for temporarily storing data close at hand rather than on separate memory chips. Compilers, which translate a computer program's high-level commands into elementary chip-based instructions, face an easier task because they have fewer choices to reconcile. Now, several companies are gambling that this type of architecture, an academic curiosity only a few years ago, will lead to faster, cheaper computers. Some commercial products with RISClike features are already available (SN: 3/12/83, p. 169).

The Hewlett-Packard Co. of Palo Alto, Calif., is making one of the largest corporate commitments to RISC architectures. In its Spectrum project, company researchers carefully measured the behavior of a wide range of computer programs to find which instructions came up most often and which functions were most useful. These studies led to a design for a family of computers that share a simple core and can easily be extended for particular applications. The company now has about 100 working prototypes and plans to introduce its first products next year.

Taking a slightly different approach, MIPS Computers Inc. in Mountain View, Calif., is designing a computer processor that allows several instructions to be executed at the same time without clogging the system. Based on research originally done at Stanford University, this RISC processor requires special compiler software to speed it up. The firm intends to sell processor "boards" that can be built into computers produced by other manufacturers.

Meanwhile, David A. Patterson, one of the earliest advocates of a simple approach to computer architecture, and his group at the University of California at Berkeley have in the last three years designed and fabricated three experimental RISC chips that, according to a variety of tests, really do run faster than more sophisticated computers. The first, RISC I, had only 31 instructions compared with the 304 instructions embedded in a VAX-11/780 superminicomputer. Their latest effort involves a RISC chip tailored for a programming language called Smalltalk. Initial results show that Smalltalk programs would run faster with this chip than on the best computer now available for running the software.

Despite the successes, some computer engineers are not convinced that running RISCs is the way to go. Such computers often require larger programs for a particular application. Manufacturers are also more familiar with the kinds of microprocessors now available and are resistant to change. Detailed, critical examinations of RISC concepts appear in this month's IEEE SPECTRUM and next month's IEEE COMPUTER.